feat(examples): add remote custom encoder orchestrator

[furionw]

Depends on #13293.

Why

Show the application-owned path enabled by #13293 without introducing a workflow product: a bespoke Dynamo worker runs a custom encoder inline, attaches its output to GenerateRequest.encoder_result, and invokes a remote stock aggregated dynamo.vllm worker.

What Change

• Add a replaceable inline VisionEncoderBackend runner inside a bespoke orchestrator worker.
• Package ordered CPU embeddings with encode_request_plane_tensor and the versioned ExternalEncoderResult contract.
• Call the remote Generate endpoint through LLMUnaryClient and serve the completed response with serve_unary_endpoint.
• Add a launch script, usage documentation, and focused orchestration tests.

flowchart LR
    Client --> Frontend
    Frontend --> Orchestrator[bespoke orchestrator worker]
    Orchestrator --> Encoder[inline custom encoder]
    Encoder -->|encoder_result| Orchestrator
    Orchestrator -->|GenerateRequest| VLLM[remote stock dynamo.vllm]
    VLLM -->|token stream| Orchestrator
    Orchestrator -->|unary completion| Frontend

Loading

Test Plan

• Applicable pre-commit hooks, Python compilation, and shell syntax checks pass.
• On GB200 job 2196083 with nvcr.io/nvidia/ai-dynamo/vllm-runtime-nightly:latest (ai-dynamo 1.5.0.dev20260914, vLLM 0.28.0), the three focused test files pass: 24 passed in 4.66s.
• Direct stock-decoder baseline: HTTP 200 and ready.
• Frontend → orchestrator/inline encoder → remote stock vLLM: HTTP 200 and The answer ... is 42.
• Debug evidence for the same feature request records handlers._assemble_external_encoder_prompt: prepared external encoder prompt; no fatal service errors were found.